Microsoft announces AMD EPYC Azure HPC VMs

Gideon

Golden Member
Nov 27, 2007
1,608
3,573
136
https://azure.microsoft.com/en-us/blog/introducing-the-new-hb-and-hc-azure-vm-sizes-for-hpc/
First, we are introducing HB-series VMs optimized for applications driven by memory bandwidth, such as fluid dynamics, explicit finite element analysis, and weather modeling. HB VMs feature 60 AMD EPYC 7551 processor cores, 4 GB of RAM per CPU core, and no hyperthreading. The AMD EPYC platform provides more than 260 GB/sec of memory bandwidth, which is 33 percent faster than x86 alternatives and 2.5x faster than what most HPC customers have in their datacenters today.

I still remember some guys saying here that EPYC is useless in HPC, despite more BW, seriously claiming that can be no viable use-cases that are memory bandwidth bound :p

I wonder why they disabled SMT on both Intel and AMD though
 

NostaSeronx

Diamond Member
Sep 18, 2011
3,683
1,218
136
Technically, out of AMD's completed designs AMD's Zen isn't the best for HPC. That would go to Zen2/EPYC2.

Disabled SMT => Microsoft's Meltdown/Spectre patches slow it down.
 
Last edited:

Cali3350

Member
May 31, 2004
127
11
81
Technically, out of AMD's completed designs AMD's Zen isn't the best for HPC. That would go to Zen2/EPYC2.

Disabled SMT => Microsoft's Meltdown/Spectre patches slow it down.
Surely that wouldnt impact Linux VM's, and Microsoft could provision dedicated Linux / Windows hosts?

Edit: Totally possible the Linux patches would starve enough bandwidth to make it impractical there too, just speaking about the Microsoft patch mentioned specifically.
 
Last edited:

Tuna-Fish

Golden Member
Mar 4, 2011
1,324
1,462
136
Technically, out of AMD's completed designs AMD's Zen isn't the best for HPC. That would go to Zen2/EPYC2.

Disabled SMT => Microsoft's Meltdown/Spectre patches slow it down.

HPC outfits don't disable SMT. Neither do they run the worst meltdown/spectre patches. HPC has complete control over all the software running on their machines and don't need to be open to internet, so they can choose performance over security.
 

Vattila

Senior member
Oct 22, 2004
799
1,351
136
HPC outfits don't disable SMT. Neither do they run the worst meltdown/spectre patches. HPC has complete control over all the software running on their machines

This discussion thread is about VM (virtual machines) for HPC offered by the Microsoft Azure cloud service. As I understand it, the patches are necessary to protect VMs from each other. Don't you consider Microsoft Azure a "HPC outfit"?
 

Vattila

Senior member
Oct 22, 2004
799
1,351
136
I wonder why they disabled SMT on both Intel and AMD though

I do too. If there is no basis for disabling SMT on AMD hardware (which is what I have read on the matter and understand to be the case), then I hope AMD is on the case to have it enabled, so that customers are not needlessly penalised, and so that there is no misunderstanding in the industry.
 

Abwx

Lifer
Apr 2, 2011
10,853
3,298
136
If a core ST perf is 100 and 140 with SMT then this latter mode will provide a throughput of only 70 for each thread, and hence per VM, not counting that using one core/VM is a way to not execute another VM in the same core with codes that are not related or even belonging to different clients..
 

moinmoin

Diamond Member
Jun 1, 2017
4,934
7,619
136
Yeah, if the customer pays per core anyway SMT being disabled is actually an advantage. Higher, more reliably reproducable performance, more cache, more memory bandwidth etc. per core/thread.
 
  • Like
Reactions: ryan20fun and IEC

Abwx

Lifer
Apr 2, 2011
10,853
3,298
136
Oracle released a benchmark on costs and perfs for these kind of apps.

https://www.computerbase.de/2018-10/oracle-cloud-amd-epyc-e2-platform/#bilder

1-1080.3065899811.jpg


2-1080.787806037.jpg

3-1080.2501966215.jpg
 

TheGiant

Senior member
Jun 12, 2017
748
353
106
So when I have average 10 hours per day calculations and I need 256GB of RAM and 32 Cores (don't see other combo) the 10 hours cost is 32*0,03*10=9,6 USD/10h ?
 

Atari2600

Golden Member
Nov 22, 2016
1,409
1,655
136
I wonder why they disabled SMT on both Intel and AMD though

Because memory bandwidth is the choke point.

Better to maximise all your bandwidth on your full fat cores than split it between 1 & 1/3rd cores.

Its something we'd have done when running fluent in the past.

Although, I suppose should also consider that licensing is per thread. So when you are paying, oh, off the top of my head, 1-2k /year /thread for the privilege, it also makes economic sense to run with SMT off.
 
Last edited:

Abwx

Lifer
Apr 2, 2011
10,853
3,298
136
What is of interest in those slides is that there s clock/clock and core/core comparison of Integer and FP perfs between Zen 1 and SKL.

Zen has 10% better INT perf and slightly hedge SKL in FP, albeit for the latter case Zen has a huge advantage in perf/watt.